GTCCT: Geography and Time-based Cloud Classification Transformer for Himawari-8 imagery

Acquiring cloud type information is crucial for analyzing atmospheric dynamics and predicting climate change. However, due to the complexity and variability of clouds, accurate cloud classification is still a challenging task. The most used cloud classification methods typically rely on the spectral indexes analysis which is not suitable for clouds that have similar spectral features. Some deep learning-based methods achieve better performance, but they only extract features from image ignoring temporal and geography characteristics. To address the above challenges, this study introduces a cloud classification method based on geography and time encoding transformer network (GTCCT). The proposed GTCCT method integrates the channel attention mechanism (CAM), as well as temporal and geographical encoding into transformer network, effectively leveraging the spatial and spectral features of Himawari-8 satellite images. Experimental validation demonstrates that the network successfully classifies seven cloud types (Ci, As, Ac, Sc, Cu, Ns, Dc) and clear sky conditions in part of Southeast Asia. Quantitative results with the CloudSat 2B-CLDCLASS product as reference show that the proposed GTCCT method achieves higher overall accuracy (82.13% versus 76.61%) and Kappa (0.7306 versus 0.6738) than baseline methods. The ablation experiment results also indicate that features related to cloud types, time, geographic and fusion of spatial and channel attention can help improve cloud classification performance of the proposed GTCCT.